Integer motion compensation
阅读说明：本技术 整数运动补偿 (Integer motion compensation ) 是由 刘鸿彬 张莉 张凯 王悦 于 2019-06-10 设计创作，主要内容包括：一种解码包括视频的数字表示的比特流的方法,包括：从比特流对当前视频块的运动信息进行解码,生成一个或多个模板,其中一个或多个模板中的每个包含具有多个样本的视频块；基于一个或多个模板的模板来细化当前视频块的运动信息；以及对细化的运动信息执行运动补偿。(A method of decoding a bitstream comprising a digital representation of video, comprising: decoding motion information of a current video block from a bitstream, generating one or more templates, wherein each of the one or more templates contains a video block having a plurality of samples; refining motion information of the current video block based on a template of the one or more templates; and performing motion compensation on the refined motion information.)
1. A method of decoding a bitstream comprising a digital representation of video, comprising:
motion information of the current video block is decoded from the bitstream,
Generating one or more templates, wherein each of the one or more templates contains a video block having a plurality of samples;
Refining motion information of the current video block based on a template of the one or more templates; and
Performing motion compensation on the refined motion information.
2. The method of claim 1, wherein the generating one or more templates comprises using one or more of a bilinear interpolation filter, integer-only motion vector values, integer-only horizontal motion vector values, and integer-only vertical motion vector values.
3. The method of claim 2, wherein the integer-only horizontal motion vector value and the integer-only vertical motion vector value are for a first prediction direction and a second prediction direction.
4. The method of claim 2, wherein the integer-only motion vector values are for only the first prediction direction.
5. The method of claim 2, wherein:
One of the integer-only horizontal motion vector value and the integer-only vertical motion vector value is for a first prediction direction; and is
One of the integer-only horizontal motion vector value and the integer-only vertical motion vector value is used for a second prediction direction.
6. The method of claim 2, wherein:
Said integer-only motion vector values are for a first prediction direction; and is
One of the integer-only horizontal motion vector value and the integer-only vertical motion vector value is used for a second prediction direction.
7. The method of claim 2, wherein one of the integer-only horizontal motion vector value and the integer-only vertical motion vector value is used for only the first prediction direction.
8. The method of claim 2, wherein the conditions comprise dimensions of the one or more templates.
9. The method of claim 8, wherein the size of the one or more templates corresponds to a smallest possible block size.
10. The method of claim 8, wherein the size of the one or more templates corresponds to a maximum possible block size.
11. The method of claim 8, wherein the size of the one or more templates corresponds to the selected block size.
12. The method of claim 2, wherein a condition comprises a shape of the one or more templates.
13. The method of claim 1, wherein the performing motion compensation on the refined motion information comprises using a longer tap filter.
14. The method of claim 13, wherein the longer tap filter is an 8-tap interpolation filter.
15. The method of claim 2, wherein the integer-only motion vector value imv is calculated from fractional motion vector values fmv as one of:
i. imv＝(fmv+(1<<(prec-1)))>>prec
ii. imv＝fmv>>prec
iii. imv＝(fmv+sign(fmv)*(1<<(prec-1)))>>prec
Where prec represents the fractional precision of the motion vector and the sign () function is defined as:
16. A video decoding apparatus, comprising: a processor configured to implement the method of one or more of claims 1 to 15.
17. A video encoding device, comprising: a processor configured to implement the method of one or more of claims 1 to 15.
18. A computer program product having stored thereon computer code which, when executed by a processor, causes the processor to carry out the method of one or more of claims 1 to 15.
19. A method, apparatus or system as described in this document.
Technical Field
This document relates to video coding techniques.
Background
Despite advances in video compression technology, digital video still accounts for the greatest bandwidth usage on the internet and other digital communication networks. As the number of connected user devices capable of receiving and displaying video increases, the bandwidth required to pre-count the usage of digital video will continue to grow.
Disclosure of Invention
Techniques related to decoder-Side Motion Vector Derivation (DMVD) in video coding are disclosed. The technique may be applied to existing Video Coding standards such as HEVC, or finalized standards (Versatile Video Coding (VVC)). The technique may also be applied to future video coding standards or video codecs.
In one example aspect, a method of decoding a bitstream comprising a digital representation of video is disclosed. The method comprises the following steps: decoding motion information of a current video block from a bitstream; generating one or more templates, wherein each of the one or more templates includes a video block having a plurality of samples, refining motion information of the current video block based on the templates of the one or more templates, and performing motion compensation on the refined motion information.
In another example aspect, an apparatus is disclosed that includes a processor configured to implement each of the above-described methods.
In yet another example aspect, the methods may be embodied in the form of computer executable instructions and stored on a computer readable program medium.
These and other aspects are further described in this document.
Drawings
Figure 1 shows an example of a derivation process for the merge (merge) candidate list construction.
Fig. 2 shows example positions of spatial Merge candidates.
Fig. 3 shows an example of a candidate pair considering redundancy check of the spatial Merge candidate.
Fig. 4A and 4B show example locations of the second PU of the N × 2N and 2N × N partitions.
Fig. 5 is an example illustration of motion vector scaling for the temporal Merge candidate.
FIG. 6 shows an example of candidate positions for the Merge candidates C0 and C1.
Fig. 7 shows an example of combined bidirectional predictive Merge candidates.
Fig. 8 shows an example derivation process of a motion vector prediction candidate.
Fig. 9 shows an example illustration of motion vector scaling for spatial motion vector candidates.
FIG. 10 illustrates an example of bilateral matching.
Fig. 11 shows an example of template matching.
Fig. 12 shows an example of one-sided ME in FRUC.
Fig. 13 shows an example of a DMVR based on double-sided template matching.
Fig. 14 shows an example of a simplified template in template matching.
Fig. 15 is a flow diagram of an example method of video decoding.
fig. 16 is a block diagram of a video decoding device.
Fig. 17 shows an example implementation of a video encoder.
Detailed Description
This document provides various techniques that may be used by a decoder of a video bitstream to improve the quality of decompressed or decoded digital video. Moreover, the video encoder may also implement these techniques during the encoding process in order to reconstruct the decoded frames for further encoding.
For ease of understanding, section headings are used in this document and do not limit embodiments and techniques to the corresponding sections. Thus, embodiments from one section may be combined with embodiments from other sections. Furthermore, while some embodiments describe the video encoding steps in detail, it should be understood that the corresponding steps of decoding undo (undo) encoding will be implemented by the decoder. Furthermore, the term video processing includes video encoding or compression, video decoding or decompression, and video transcoding, where video pixels are represented from one compression format to another compression format or at different compression bit rates.
1. Technical framework
Video coding standards have evolved primarily by developing the well-known ITU-T and ISO/IEC standards. ITU-T produces H.261 and H.263, ISO/IEC produces MPEG-1 and MPEG-4 Visual, and both organizations jointly produce H.262/MPEG-2 Video and H.264/MPEG-4 Advanced Video Coding (AVC) and H.265/HEVC standard [1 ]. Starting from h.262, video coding standards are based on hybrid video coding structures that utilize temporal prediction plus transform coding. In order to explore future Video coding techniques other than HEVC, Joint Video Exploration Team (jfet) was established by VCEG and MPEG in 2015. Since then, many new methods have been adopted by JFET and introduced into a reference software named Joint Exploration Model (JEM) [3] [4 ]. In month 4 of 2018, Joint video expert Team (jviet) between VCEG (Q6/16) and ISO/IEC JTC1 SC29/WG11(MPEG) was created for the VVC standard, aiming at a 50% reduction in bit rate compared to HEVC.
Inter prediction in HEVC/H.265
Each inter-predicted Prediction Unit (PU) has motion parameters for one or two reference picture lists. The motion parameters include a motion vector and a reference picture index. Inter _ pred _ idc may also be used to signal the use of one of the two reference picture lists. Motion vectors can be coded explicitly as deltas relative to the predictor.
When a Coding Unit (CU) is coded in skip mode, one PU is associated with the CU and there are no significant residual coefficients, no coded motion vector delta, or reference picture indices. The Merge mode is specified such that the motion parameters, including spatial and temporal candidates, of the current PU are obtained from neighboring PUs. The Merge mode may be applied to any inter-predicted PU, not just the skip mode. An alternative to the Merge mode is the explicit transmission of motion parameters, where for each PU motion vectors (more precisely, motion vector differences compared to motion vector predictors), corresponding reference picture indices and reference picture list usage for each reference picture list are explicitly signaled [2 ]. In this document, this mode is referred to as Advanced Motion Vector Prediction (AMVP).
When the signaling indicates that one of the two reference picture lists is to be used, the PU is generated from one sample block. This is called "uni-prediction". Unidirectional prediction may be used for both P-slices and B-slices.
When the signaling indicates that two reference picture lists are to be used, the PU is generated from two blocks of samples. This is called "bi-prediction". Bi-directional prediction only applies to B slices.
The following text provides detailed information of the inter prediction modes specified in HEVC. The description will start with the Merge mode.
2.1.1Merge mode
Candidate derivation for 2.1.1.1Merge mode
When predicting a PU using the Merge mode, the index pointing to an entry in the Merge candidate list is parsed from the bitstream and used to retrieve motion information. The construction (construction) of this list is specified in the HEVC standard and can be summarized according to the following sequence of steps:
● step 1: initial candidate derivation
Step 1.1: spatial candidate derivation
Step 1.2: redundancy check of spatial candidates
Step 1.3: temporal candidate derivation
● step 2: inserting additional candidates
Step 2.1: creating bi-directional prediction candidates
Step 2.2: inserting zero motion candidates
These steps are also schematically depicted in fig. 1. For spatial Merge candidate derivation, a maximum of four Merge candidates are selected among the candidates located at five different positions. For the temporal Merge candidate derivation, at most one Merge candidate is selected among the two candidates. Since a constant number of candidates is assumed at the decoder for each PU, additional candidates are generated when the number of candidates obtained from step 1 does not reach the maximum number of Merge candidates (MaxNumMergeCand) signaled in the slice header. Since the number of candidates is constant, the index of the best Merge candidate is encoded using truncated unary binarization (TU). If the size of the CU is equal to 8, all PUs of the current CU share a single Merge candidate list, which is the same as the Merge candidate list of the 2N × 2N prediction unit.
hereinafter, operations related to the above steps are described in detail.
2.1.1.2 spatial candidate derivation
In the derivation of spatial Merge candidates, a maximum of four Merge candidates are selected among the candidates located at the positions depicted in FIG. 2. The order of derivation is A1、B1、B0、A0And B2. Only when in position A1、B1、B0、A0Does not consider location B when any PU of (e.g., because it belongs to another slice or block) is unavailable or intra-coded2. At the addition position A1After the candidate of (b), a redundancy check is performed on the addition of the remaining candidates, which ensures that candidates with the same motion information are excluded from the list, resulting in an improved coding efficiency. In order to reduce computational complexity, not all possible candidate pairs are considered in the mentioned redundancy check. Instead, only the pairs connected with arrows in fig. 3 are considered and candidates are added to the list only if the corresponding candidates for redundancy check have different motion information. Another source of duplicate motion information is the "second PU" associated with a partition other than 2 nx 2N. As an example, fig. 4A and 4B depict the second PU for the N × 2N and 2N × N cases, respectively. Position A when the current PU is partitioned into Nx 2N1Is not selectedConsidered for list construction. In fact, adding this candidate will result in two prediction units with the same motion information, which is redundant for having only one PU in the coding unit. Similarly, when the current PU is partitioned into 2 NxN, position B is not considered1。
2.1.1.3 temporal candidate derivation
In this step, only one candidate is added to the list. In particular, in the derivation of the temporal Merge candidate, the scaled motion vector is derived based on a co-located PU belonging to the picture within the given reference picture list having the smallest POC difference from the current picture. The reference picture list to be used for deriving the co-located PU is explicitly signaled in the slice header. A scaled motion vector for the temporal Merge candidate is obtained as shown by the dashed line in fig. 5, scaled from the motion vector of the co-located PU using POC distances tb and td, where tb is defined as the POC difference between the reference picture of the current picture and td is defined as the POC difference between the reference picture of the co-located picture and the co-located picture. The reference picture index of the temporal Merge candidate is set equal to zero. The actual implementation of the scaling process is described in HEVC specification [1 ]. For B slices, two motion vectors are obtained, one for reference picture list0 and the other for reference picture list1, and the two motion vectors are combined to obtain the bi-predictive Merge candidate.
In co-located PU (Y) belonging to a reference frame, in candidate C0And C1The position of the time candidate is selected in between, as shown in fig. 6. If at position C0Is unavailable, intra-coded, or is outside the current CTU row, then position C is used1. Otherwise, position C0for derivation of the time Merge candidate.
2.1.1.4 insertion of additional candidates
In addition to spatial and temporal Merge candidates, there are two additional types of Merge candidates: a combined bi-directional predicted Merge candidate and zero Merge candidate. A combined bi-directional predicted Merge candidate is generated by using spatial and temporal Merge candidates. The combined bi-directionally predicted Merge candidates are for B slices only. A combined bi-directional prediction candidate is generated by combining the first reference picture list motion parameters of the initial candidate with the second reference picture list motion parameters of the other candidate. If these two tuples provide different motion hypotheses, they will form a new bi-directional prediction candidate. As an example, fig. 7 depicts the case when two candidates with mvL0 and refIdxL0 or mvL1 and refIdxL1 in the original list (left side) are used to create a combined bi-predictive Merge candidate that is added to the final list (right side). A number of rules are defined in the HEVC specification regarding combinations that are considered to generate these additional Merge candidates.
Zero motion candidates are inserted to fill the remaining entries in the Merge candidate list to reach MaxUMMergeCand capacity. These candidates have zero spatial displacement and a reference picture index that starts from zero and increases each time a new zero motion candidate is added to the list. The number of reference frames that these candidates use is 1 and 2, for unidirectional and bi-directional prediction, respectively. Finally, no redundancy check is performed on these candidates.
2.1.1.5 motion estimation regions for parallel processing
To speed up the encoding process, motion estimation may be performed in parallel, thereby deriving motion vectors for all prediction units within a given region simultaneously. Deriving the Merge candidate from the spatial neighborhood may interfere with parallel processing because one prediction unit cannot derive motion parameters from neighboring PUs until its associated motion estimation is complete. To mitigate the trade-off between coding efficiency and processing latency, HEVC defines a Motion Estimation Region (MER), the size of which is signaled in the picture parameter set using the syntax element "log 2_ parallel _ merge _ level _ minus 2" of HEVC. When MER is defined, the Merge candidates falling in the same region are marked as unavailable and are therefore not considered in the list construction.
2.1.2AMVP
AMVP exploits the spatial-temporal correlation of motion vectors to neighboring PUs, which is used for explicit transmission of motion parameters. For each reference picture list, a motion vector candidate list is constructed by first checking the availability of the top, left side, of temporally neighboring PU locations, removing redundant candidates and adding zero vectors to make the candidate list a constant length. The encoder may then select the best predictor from the candidate list and send a corresponding index indicating the selected candidate. Similar to the Merge index signaling, the index of the best motion vector candidate is coded using a truncated unary. The maximum value to be encoded in this case is 2 (see fig. 8). In the following sections, details are provided regarding the derivation process of motion vector prediction candidates.
2.1.2.1 derivation of AMVP candidates
Fig. 8 summarizes the derivation of motion vector prediction candidates.
In motion vector prediction, two types of motion vector candidates are considered: spatial motion vector candidates and temporal motion vector candidates. For spatial motion vector candidate derivation, two motion vector candidates are finally derived based on the motion vectors of each PU located at five different positions as shown in fig. 2.
For temporal motion vector candidate derivation, one motion vector candidate is selected from two candidates, which are derived based on two different co-located positions. After generating the first list of spatio-temporal candidates, duplicate motion vector candidates in the list are removed. If the number of potential candidates is greater than 2, the motion vector candidate whose reference picture index is greater than 1 within the associated reference picture list is removed from the list. If the number of spatio-temporal motion vector candidates is less than 2, additional zero motion vector candidates are added to the list.
2.1.2.2 spatial motion vector candidates
In the derivation of spatial motion vector candidates, a maximum of two candidates are considered among the five potential candidates, which are derived from PUs located at the positions shown in fig. 2, those positions being the same as the position of the motion Merge. Defining the derivation order of the left side of the current PU as A0、A1And scaled A0Zoom of A1. Defining the derivation order of the upper side of the current PU as B0、B1、B2Scaled B0Zoomed B1Zoomed B2. Thus, it is possible to provideFor each side, there are four cases that can be used as motion vector candidates, two of which do not require the use of spatial scaling and two of which use spatial scaling. Four different cases are summarized below:
● no spatial scaling
- (1) identical reference picture list, and identical reference picture index (identical POC)
- (2) different reference picture lists, but the same reference picture (same POC)
● spatial scaling
- (3) same reference picture list, but different reference pictures (different POCs)
- (4) different reference picture lists, and different reference pictures (different POCs)
The case of no spatial scaling is checked first and then the spatial scaling is checked. Spatial scaling is considered when POC differs between reference pictures of neighboring PUs and reference pictures of a current PU regardless of reference picture lists. If all PUs of the left candidate are not available or intra coded, scaling of the motion vectors described above is allowed to facilitate parallel derivation of the left and upper MV candidates. Otherwise, spatial scaling of the motion vectors is not allowed.
In the spatial scaling process, the motion vectors of neighboring PUs are scaled in a similar manner as in the temporal scaling, as shown in fig. 9. The main difference is that the reference picture list and the index of the current PU are given as input; the actual scaling procedure is the same as the time scaling procedure.
2.1.2.3 temporal motion vector candidates
All processes for deriving the temporal Merge candidate are the same as those for deriving the spatial motion vector candidate, except for the reference picture index derivation (see FIG. 6). The reference picture index is signaled to the decoder.
2.2 novel inter-frame prediction method in JEM
2.2.1 motion vector derivation for pattern matching
The Pattern-Matched Motion Vector Derivation (PMMVD) mode is a special Merge mode based on Frame-Rate Up Conversion (FRUC) techniques. With this mode, instead of signaling the motion information of the block, the motion information of the block is derived at the decoder side.
When the Merge flag of the CU is true, a FRUC flag is signaled to the CU. When the FRUC flag is false, the Merge index is signaled and the normal Merge mode is used. When the FRUC flag is true, an additional FRUC mode flag is signaled to indicate which method (bilateral matching or template matching) will be used to derive the motion information for the block.
At the encoder side, the decision on whether to use frucemerge mode for a CU is based on RD cost selection for normal Merge candidates. That is, both of the two matching patterns (bilateral matching and template matching) of a CU are verified by using RD cost selection. The matching pattern that results in the least cost is further compared to other CU patterns. If the FRUC matching pattern is the most efficient pattern, the FRUC flag is set to true for the CU and the associated matching pattern is used.
The motion derivation process in frucemarge mode has two steps: CU-level motion search is performed first, followed by sub-CU-level motion refinement. At the CU level, an initial motion vector for the entire CU is derived based on bilateral matching or template matching. First, a list of MV candidates is generated and the candidate that results in the smallest matching cost is selected as the starting point for further CU-level refinement. Then, local search based on bilateral matching or template matching is performed near the start point, and the MV result of the minimum matching cost is taken as the MV of the entire CU. Subsequently, the motion information is further refined at sub-CU level, taking the derived CU motion vector as a starting point.
For example, the following derivation process is performed for W × HCU motion information derivation. In the first stage, the MVs for the entire W × HCU are derived. In the second stage, the CU is further divided into M × M sub-CUs. The value of M is calculated as shown in (1), D is a predefined division depth, and is set to 3 by default in JEM. The MV of each sub-CU is then derived.
As shown in fig. 10, motion information of a current CU is derived using bilateral matching by finding a closest match between two blocks along a motion trajectory of the current CU in two different reference pictures. Under the assumption of a continuous motion trajectory, the motion vectors MV0 and MV1 pointing to the two reference blocks should be proportional to the temporal distance between the current picture and the two reference pictures (i.e., TD0 and TD 1). As a special case, the bilateral matching becomes a mirror-based bi-directional MV when the current picture is temporally between two reference pictures and the temporal distance from the current picture to the two reference pictures is the same.
As shown in fig. 11, template matching is used to derive motion information for a current CU by finding the closest match between a template (the top and/or left neighboring blocks of the current CU) in the current picture and a block (of the same size as the template) in a reference picture. In addition to the FRUCMerge pattern described above, template matching is also applied to the AMVP pattern. In JEM, AMVP has two candidates, as is done in HEVC. New candidates are derived by the template matching method. If the newly derived candidate by template matching is different from the first existing AMVP candidate, it is inserted at the very beginning of the AMVP candidate list, and then the list size is set to 2 (meaning the second existing AMVP candidate is removed). When applied to AMVP mode, only CU level search is applied.
2.2.2 CU-LEVEL MV candidate sets
MV candidates set at the CU level include:
(i) If the current CU is in AMVP mode, it is the original AMVP candidate
(ii) All of the large candidates are selected from the group,
(iii) Several MVs in the interpolated MV field introduced in section 2.2.4.
(iv) Top and left adjacent motion vectors
When using bilateral matching, each valid MV of the Merge candidate is used as an input to generate a MV pair assuming bilateral matching. For example, in reference list a, one valid MV of the Merge candidate is (MVa, refa). Then, the reference picture refb of its paired bilateral MV is found in the other reference list B, so that refa and refb are temporally located on different sides of the current picture. If such refb is not available in reference list B, refb is determined to be a different reference from refa and its temporal distance to the current picture is the minimum in list B. After determining refb, MVb is derived by scaling MVa based on the temporal distance between the current picture refa and refb.
Four MVs from the interpolated MV field are also added to the CU level candidate list. More specifically, interpolation MVs at positions (0,0), (W/2,0), (0, H/2), and (W/2, H/2) of the current CU are added.
when FRUC is applied to AMVP mode, the original AMVP candidate is also added to the CU-level MV candidate set.
At the CU level, a maximum of 15 MVs are added to the candidate list for AMVP CUs and a maximum of 13 MVs are added to the candidate list for MergeCU.
2.2.3 sub-CU level MV candidate set
The MV candidates set at the sub-CU level include:
(i) the determined MV is searched from the CU level,
(ii) Top, left side, top left corner and top right corner adjacent MVs,
(iii) A scaled version of the collocated MV from the reference picture,
(iv) A maximum of 4 ATMVP candidates,
(v) A maximum of 4 STMVP candidates
The scaled MV from the reference picture is derived as follows. All reference pictures in both lists are traversed. The MVs at the collocated positions of the sub-CUs in the reference picture are scaled to the reference of the starting CU level MV.
ATMVP and STMVP candidates are limited to the first four.
At the sub-CU level, a maximum of 17 MVs are added to the candidate list.
2.2.4 Generation of interpolated MV fields
Before encoding a frame, an interpolation motion field is generated for the whole picture based on one-sided ME. The motion field can then be used later as a CU-level or sub-CU-level MV candidate.
First, the motion field of each reference picture in the two reference lists is traversed at 4 × 4 block level. For each 4 x 4 block, if the motion associated with the block passes through a 4 x 4 block in the current picture (as shown in fig. 12) and the block is not assigned any interpolated motion, the motion of the reference block is scaled to the current picture according to temporal distances TD0 and TD1 (in the same way as MV scaling of TMVP in HEVC) and the scaled motion is assigned to the block in the current frame. If no scaled MV are assigned to a 4 x 4 block, the motion of the block is marked as unavailable in the interpolated motion field.
2.2.5 interpolation and matching costs
When the motion vector points to a fractional sample position, motion compensated interpolation is required. To reduce complexity, instead of conventional 8-tap HEVC interpolation, bilinear interpolation is used for both edge matching and template matching.
The computation of the matching cost is somewhat different at different steps. When selecting candidates from the CU-level candidate set, the matching cost is the Absolute Sum Difference (SAD) of the bilateral matching or template matching. After determining the starting MV, the matching cost for the bilateral matching of the sub-CU level search is calculated as follows:
where w is a weighting factor empirically set to 4, MV and MVsIndicating the current MV and the starting MV, respectively. SAD is still used as the matching cost for template matching for sub-CU level search.
In FRUC mode, the MV is derived by using only the luma samples. The derived motion will be used for both luma and chroma for MC inter prediction. After the MV is determined, the final MC is performed using an 8-tap interpolation filter for luminance and a 4-tap interpolation filter for chrominance.
2.2.6 MV refinement
MV refinement is a pattern-based MV search, with a bilateral matching cost or template matching cost as criteria. In JEM, two Search modes are supported-the unconstrained Center-Biased Diamond Search (UCBDS) and the adaptive Cross Search, with MV refinement at the CU level and sub-CU level, respectively. For both CU and sub-CU level MV refinement, the MV is searched directly with quarter luma sample MV precision, and then one-eighth luma sample MV refinement. The search range for MV refinement for the CU and sub-CU steps is set equal to 8 luma samples.
2.2.7 selection of prediction Direction in template matching FRUCMerge mode
In the bilateral matching Merge mode, bi-prediction is always applied, since the motion information of a CU is derived based on the closest match between two blocks along the motion trajectory of the current CU in two different reference pictures. There is no such restriction on template matching Merge patterns. In the template matching Merge mode, the encoder may select a CU from among unidirectional prediction in list0, unidirectional prediction in list1, or bidirectional prediction. The selection is based on the following template matching costs:
If costBi & gt factor & ltmin (cost0, cost1)
Then bi-directional prediction is used;
Otherwise, if cost0< ═ cost1
Then the one-way prediction in list0 is used;
If not, then,
Using the unidirectional prediction in table 1;
Where cost0 is the SAD of the List0 template match, cost1 is the SAD of the List1 template match, and cost Bi is the SAD of the bidirectional prediction template match. The value of factor is equal to 1.25, which means that the selection process is biased towards bi-directional prediction.
Inter prediction direction selection is only applied to the CU-level template matching process.
2.2.8 decoder-side motion vector refinement
In the bi-directional prediction operation, for prediction of one block region, two prediction blocks formed using a Motion Vector (MV) of list0 and an MV of list1, respectively, are combined to form a single prediction signal. In the Decoder-Side Motion Vector Refinement (DMVR) method, two Motion vectors for bi-directional prediction are further refined by a two-sided template matching process. The bilateral template matching is applied in the decoder to perform a distortion-based search between the bilateral template and reconstructed samples in the reference picture in order to obtain refined MVs without the need to transmit additional motion information.
In DMVR, the two-sided template is generated as a weighted combination (i.e., average) of the two prediction blocks from the initial MV0 of list0 and MV1 of list1, respectively, as shown in fig. 13. The template matching operation includes calculating a cost metric between the generated template and a sample region (around the initial prediction block) in the reference picture. For each of the two reference pictures, the MV yielding the smallest template cost is considered as the updated MV of the list to replace the original MV. In JEM, nine MV candidates are searched for each list. The nine MV candidates include the original MV and 8 surrounding MVs with a luma sample offset in the horizontal or vertical direction or both directions from the original MV. Finally, two new MVs, MV0 'and MV1', as shown in fig. 13, are used to generate the final bi-directional prediction result. The Sum of Absolute Differences (SAD) is used as a cost measure. Note that when calculating the cost of a prediction block generated by one surrounding MV, the prediction block is actually obtained using rounded MVs (to integer pixels) instead of real MVs.
DMVR is applied to the bidirectionally predicted Merge mode, where one MV is from a past reference picture and another MV is from a future reference picture, without transmitting additional syntax elements. In JEM, DMVR is not applied when LIC, affine motion, FRUC, or sub-cumarge candidates are enabled for a CU.
2.2.9 example of the problem
DMVD methods such as DMVR and FRUC perform motion estimation to derive or refine motion information, which is very complex for the decoder. During motion estimation, they have a common problem: the differences (absolute differences, squared differences, etc.) between the template and candidate blocks are calculated and summed for all pixels in the block and then used to select the best matching block. This is not necessary as the difference of the partial pixels may be sufficient to select the best candidate block or MV. Meanwhile, only the luminance component is generally used in the derivation or refinement of the motion vector, without considering the chrominance component.
For DMVR, it has another complexity problem: it performs motion compensation twice, once for generating the template and once for generating the final prediction block. As a result, for each reference picture list (i.e., prediction direction), it performs both horizontal interpolation and vertical interpolation twice, in case the initial MV and refined MV have only fractional components. This greatly increases the worst case complexity. Meanwhile, the DMVR operates only in the Merge mode and cannot operate in the AMVP mode. In MV refinement, it takes the signaled MV (MV derived from the Merge candidate) as the starting MV and checks its surrounding MVs. However, the MV accuracy of the signaled MV is not considered. In AMVR, it is possible to select a low precision MV. For example, assuming that the highest allowed MV precision is 1/4 pixels, in AMVR, either a 4-pixel or 1-pixel MV may be used. In this case, DMVR can be used to refine MV precision. Unlike FRUC, which can be applied at the sub-block level, DMVR executes at the block level, except for ATMVP and STMVP cases, which can result in coding performance loss.
For the FURC, when performing bilateral matching, the MV difference between the starting MV and the candidate MV is considered to suppress unreliable motion vectors, as in equation 2. This may not be justified by multiplying the MV differences by a fixed weighting factor. For larger blocks, SAD plays a dominant role and MV difference is negligible, and for smaller blocks MV difference may be too large.
2.2.10 example embodiment
We propose several aspects to reduce the complexity of DMVD methods and improve coding performance. The disclosed method can be applied to existing DMVD methods, but also to future methods of motion/mode derivation at the decoder side.
First, the cost between the template and the candidate block is calculated only for a portion of the pixels in the decoder-side motion estimation, i.e., in the motion information derivation or refinement process (e.g., taking into account the difference of distortion and MV, distortion, or cost). Second, for DMVR, the number of interpolations is reduced. Third, DMVR is applied to AMVP mode using some embodiments of the disclosed technology. Fourth, the weighting factor for the MV difference may be different for different block sizes.
The examples listed below provide some ways in which the disclosed techniques may be embodied in a video decoding process.
The prec is expressed as motion vector precision, which means that the motion vector has 1/2^ N pixel precision when the prec equals N. N may be a positive integer, zero, or a negative integer.
1. The cost (e.g., difference) between the template and the candidate block is computed only for partially selected lines in the motion information derivation or refinement process.
a. In one example, the selected row is defined as the ith row of all every N rows, where N >1 and 1< ═ i < ═ N. For example, N equals 2 and i equals 1.
b. In one example, for each group having N rows, some of the rows within the group are used as the selected rows. For example, a first row and a second row of every 4 rows are utilized.
c. In one example, the cost is calculated for any selected row of the block, e.g., the first and last rows, or the first two rows and the last two rows.
d. the same rule may be applied to all block sizes when selecting partial rows. Alternatively, different rules may be applied for different block sizes and/or block shapes (e.g., squares or rectangles or the ratio between block width and block height).
i. in one example, during cost calculation, for larger block sizes, more rows are skipped, and vice versa. For example, when the block size is less than 16 × 16 (i.e., width x height <16 x 16), the difference is calculated for the first row of every 2 rows, but for other block sizes the difference is calculated for the first row of every 4 rows.
in one example, during cost calculation, for block shapes with greater height, more rows are skipped, and vice versa. For example, when the height of a block is less than 16, the cost is calculated for the first row of every 2 rows, but for other block sizes, the cost is calculated for the first row of every 4 rows.
in one example, this simplification applies only to one or a few minimum block sizes (i.e., minimum width x height) to suppress worst-case complexity. For example, the simplification applies only to blocks with an area smaller than 8 × 8.
in one example, this simplification only applies to one or a few maximum block sizes. For example, the simplification applies only to blocks with an area larger than 32 × 32.
v. in one example, this simplification applies only to one or a few block shapes with the largest block height or width.
In one example, this simplification applies to only some selected block shapes.
2. For each row of the block or each selected row of the block, the cost of all or only a portion of the columns is calculated.
a. In one example, the cost of M consecutive columns (which may start at any active column Y) per T column is calculated, where T >0, 1< ═ M < ═ T, 1< ═ Y < ═ T-M + 1. For example, T ═ 8, M ═ 4, and Y ═ 1.
b. In one example, the cost of M selected columns per T columns is calculated.
c. In one example, the cost of M arbitrarily selected columns of rows (e.g., the first K columns and the last L columns) is calculated.
d. The same rule may be applied to all block sizes when selecting partial columns. Alternatively, different rules may be applied to different block sizes and/or block shapes (e.g., squares or rectangles or the ratio between block width and block height).
i. In one example, during cost calculation, more columns are skipped for larger block sizes and vice versa. For example, when the block size is smaller than 16 × 16, the difference of the first 4 columns of each 8 columns is calculated, but the difference of the first 4 columns of each 16 columns is calculated for other block sizes. When the columns are less than 8 or 16, only the first 4 columns are used to calculate the difference.
in one example, during cost calculation, more columns are skipped for block shapes with larger widths, and vice versa. For example, when the width of the block is less than 16, the cost of the first 4 columns of every 8 columns is calculated, but the cost of the first 4 columns of every 16 columns is calculated for other block sizes.
in one example, this simplification applies only to one or a few minimum block sizes to suppress worst-case complexity.
in one example, this simplification only applies to one or a few maximum block sizes.
v. in one example, this simplification applies only to one or a few block shapes with the largest block width.
In one example, this simplification applies to only some selected block shapes.
3. In DMVR, when generating a template, motion compensation is performed using MVs or integer MVs having integer horizontal or vertical components, instead of using real MVs as in JEM.
a. in one example, for both prediction directions, the MV (both horizontal and vertical components) is rounded to integer precision.
b. In one example, the MVs of one prediction direction are rounded to integer precision, while the MVs of the other prediction directions do not change.
c. In one example, only one MV component (horizontal or vertical) is rounded to integer precision for each prediction direction.
d. in one example, the MV of one prediction direction is rounded to integer precision and only one MV component of the other prediction direction is rounded to integer precision.
e. In one example, the MV for one prediction direction is not changed and only one MV component for the other prediction direction is rounded to integer precision.
f. Fmv is represented as a fractional mv, and imv is represented as a rounded integer precision mv. Sign (x) is denoted as sign of x, an
i.imv＝(fmv+(1＜＜(prec-1)))＞＞prec
Alternatively, imv ═ fmv > prec
Alternatively, imv ═ fmv + sign (fmv) (1 < (prec-1))) > prec
g. This simplification may apply to all block sizes or only to one or a few block sizes and/or some block shapes.
i. In one example, it is applied to one or several minimum block sizes, such as 4 × 4 in JEM or BMS (reference set), or 4 × 8 and 8 × 4 in HEVC.
in one example, it is applied to one or a few maximum block sizes.
in one example, it is applied to certain selected block sizes.
4. Alternatively, in DMVR, a shorter-tap interpolation filter (such as a bilinear filter) is used in motion compensation when generating the template.
5. DMVR is proposed to be performed at a sub-block level. The block may be divided into sub-blocks in different ways.
a. In one example, all blocks are divided into fixed M × N sub-block sizes, e.g., 4 × 4, or 4 × 8, or 8 × 4, or 8 × 8, or 8 × 16, or 16 × 8, or 16 × 16, etc. When the block width/height is an integer multiple of the sub-block width/height, it is divided into sub-blocks; otherwise, it is not divided into sub-blocks.
b. in one example, a block is divided into K sub-blocks of equal size, where K > -2. For example, the M × N block is divided into 4 (M/2) × (N/2) sub-blocks, or 2 (M/2) × N sub-blocks, or 2M × (N/2) blocks.
c. In one example, the partitioning method depends on block size or block shape or other coding information. For example, an 8 × 32 block is divided into 4 × 8 sub-blocks, and a 32 × 8 block is divided into 8 × 4 sub-blocks.
d. In one example, the derived motion information for the entire block may be utilized when generating the template for the sub-block, as in the current block level DMVR.
i. Alternatively, the refined motion information of the neighboring sub-block(s) with or without the derived motion information of the entire block may be utilized to form the template.
e. In one example, the search points of the sub-blocks may also take into account refined motion information from other sub-block(s).
6. In one example, the template used for template matching (in PMMVD) includes only pixels above the current block and does not include pixels to the left of the current block, as shown in fig. 14.
7. In the existing DMVD method, a motion vector is derived or refined considering only a luminance component. The proposal also takes into account the chrominance component. The cost of the three color components of a given motion vector is denoted by Ci (where i indicates the color component index).
a. The final cost is defined as Wi Ci, where Wi indicates the weight of the ith color component.
b. Alternatively, the final cost is defined as (W0 × C0+ W1 — (C1+ C2)). In some examples, W0 or W1 is equal to 1.
c. in one example, when applying DMVR to chroma components, rounding of the motion vectors may be applied so that integer motion vectors may be utilized without applying interpolation to the chroma components.
d. in one example, when applying DMVR to chrominance components, a shorter-tap interpolation filter (such as a bilinear filter) may be applied if interpolation is needed.
8. The above method may be applied to some or all color components.
a. Different rules may be applied to different color components, or different rules may be utilized for the luma and chroma components.
b. Alternatively, how and whether the above method is applied may be further signaled in a sequence parameter set, a picture parameter set, a slice header, etc.
Fig. 15 is a flow diagram of an example method 1500 of video decoding. The method 1500 includes decoding (1502) motion information of a current video block from a bitstream; decoding motion information of a current video block from a bitstream, generating (1504) one or more templates, wherein each of the one or more templates comprises a video block having a plurality of samples, refining (1506) the motion information of the current video block based on the templates of the one or more templates, and performing (1508) motion compensation on the refined motion information.
Section 2.2.10 provides additional example embodiments and variations that may be implemented by the method 1500. For example, in some embodiments, the generating of the one or more templates comprises using one or more of a bilinear interpolation filter, integer-only motion vector values, integer-only horizontal motion vector values, and integer-only vertical motion vector values. In some embodiments, only integer horizontal motion vector values and only integer vertical motion vector values are used for the first prediction direction and the second prediction direction. In some embodiments, only integer motion vector values are used for the first prediction direction only.
In some embodiments, the conditions include dimensions of one or more templates. In some embodiments, the size of one or more templates corresponds to the smallest possible block size. In some embodiments, the size of one or more templates corresponds to the largest possible block size.
In some embodiments, the conditions include the shape of one or more templates.
As discussed further in section 2.2.10, in some embodiments, only one of the integer horizontal motion vector values and only the integer vertical motion vector values is used for the first prediction direction and only one of the integer horizontal motion vector values and only the integer vertical motion vector values is used for the second prediction direction. Alternatively, only integer motion vector values are used for the first prediction direction and only one of integer horizontal motion vector values and only integer vertical motion vector values are used for the second prediction direction.
FIG. 16 illustrates a block diagram of an example embodiment of a hardware device 1600 that may be used to implement portions of the presently disclosed technology. Hardware device 1600 may be a laptop, smartphone, tablet, camera, or other type of device capable of processing video. The device 1600 includes a processor or controller 1602 that processes data, and a memory 1604 that communicates with, stores, and/or buffers data with the processor 1602. For example, the processor 1602 may include a Central Processing Unit (CPU) or a Microcontroller Unit (MCU). In some embodiments, the processor 1602 may include a Field-Programmable Gate-Array (FPGA). In some embodiments, device 1600 includes or communicates with a Graphics Processing Unit (GPU), a Video Processing Unit (VPU), and/or a wireless communication Unit to implement various visual and/or communication data Processing functions of a smartphone device. For example, the memory 1604 may include and store processor executable code that, when executed by the processor 1602, configures the device 1600 to perform various operations, such as, for example, receiving information, commands, and/or data, processing information and data, and transmitting or providing processed information/data to another device, such as an actuator or an external display. To support various functions of device 1600, memory 1604 may store information and data such as instructions, software, values, images, and other data processed or referenced by processor 1602. For example, the storage function of the Memory 1604 may be implemented using various types of Random Access Memory (RAM) devices, Read Only Memory (ROM) devices, flash Memory devices, and other suitable storage media. The device 1600 may also include dedicated video processing circuitry 1606 for performing iterative computational functions, such as transform and decoding.
Fig. 17 is a block diagram illustrating an example implementation of a video encoder. A video encoder operates on video pictures that are encoded using encoding tools such as transform, motion estimation, and residual error coding. The encoded video frame is reconstructed (motion compensated) at the encoder and used as a reference picture for inter-coding of other video pictures. The techniques described in this document may be implemented by a video encoder or a video decoder using a hardware platform such as that described with respect to fig. 16.
From the foregoing it will be appreciated that specific embodiments of the presently disclosed technology have been described herein for purposes of illustration, but that various modifications may be made without deviating from the scope of the invention. Accordingly, the presently disclosed technology is not limited, except as by the appended claims.
The disclosed and other embodiments, modules, and functional operations described in this document can be implemented in digital electronic circuitry, or in computer software, firmware, or hardware, including the structures disclosed in this document and their structural equivalents, or in combinations of one or more of them. The disclosed and other embodiments can be implemented as one or more computer program products, i.e., one or more modules of computer program instructions encoded on a computer-readable medium for execution by, or to control the operation of, data processing apparatus. The computer readable medium can be a machine-readable storage device, a machine-readable storage substrate, a memory device, a composition of matter effecting a machine-readable propagated signal, or a combination of one or more of them. The term "data processing apparatus" includes all apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them. A propagated signal is an artificially generated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus.
a computer program (also known as a program, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program does not necessarily correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.
The processes and logic flows described in this document can be performed by one or more programmable processors executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).
Processors suitable for the execution of a computer program include, by way of example, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer. Generally, a processor will receive instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a processor for executing instructions and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks. However, a computer need not have such devices. Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, such as internal hard disks or removable disks; magneto-optical disks; and CD ROM and DVD-ROM disks. The processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry.
while this patent document contains many specifics, these should not be construed as limitations on the scope of any invention or of what may be claimed, but rather as descriptions of features specific to particular embodiments of particular inventions. Certain features that are described in this patent document in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various functions described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Furthermore, although features may be described as acting in certain combinations and even initially claimed as such, one or more features from a combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.
Also, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. Moreover, the separation of various system components in the embodiments described in this patent document should not be understood as requiring such separation in all embodiments.
Only a few embodiments and examples have been described and other embodiments, enhancements and variations can be made based on what is described and illustrated in this patent document.